Team  DARwIn-‐XOS  Team  Description  for  Humanoid  TeenSize  League  of  RoboCup  2012  

 Nicholas  McGill,  William  McGill,  Stephen  McGill,  Dean  Wilhelmi,  

Professor  Dan  Lee  &  Professor  Dennis  Hong    

GRASP  Lab  School  of  Engineering  and  Applied  Science,  University  of  Pennsylvania,  USA  

E-‐mail:  ddlee@seas.upenn.edu  Web:  http://www.seas.upenn.edu/~robocup/  

 RoMeLa  

Department  of  Mechanical  Engineering,  Virginia  Tech,  USA  E-‐mail:  dhong@vt.edu  

Web:  www.romela.org/robocup    Abstract:  This  paper  details  the  hardware,  software,  and  electrical  design  of  the  humanoid  robot  family,  DARwIn  (Dynamic  Anthropomorphic  Robot  with  Intelligence)  –  a  robot  family  designed  as  a  platform  for  researching  bipedal  motions.  The  DARwIn  family  was  the  first  US  entry  into  the  humanoid  division  of  RoboCup;  this  year,  we  hope  to  have  our  first  TeenSize  competitor.   1  Introduction  The  DARwIn  (Dynamic  Anthropomorphic  Robot  with  Intelligence)  series  robot  is  a  family  of  humanoid  robots  capable  of  bipedal  walking  and  performing  human-‐like  motions   (Fig.   1).   DARwIn   is   a   research   platform   developed   at   the   Robotics   and  Mechanisms  Laboratory   (RoMeLa)   at   Virginia  Tech   for   studying   robot   locomotion  and  sensing.  It  was  also  utilized  as  the  base  platform  for  Virginia  Tech’s  first  entry  to  the  humanoid  division  of  RoboCup  2007.  [1,2].  The  101  cm  tall,  4  kg  DARwIn-‐XOS  (a  branch  of   the  DARwIn  platform)  has  20  degrees-‐of-‐freedom  (DOF)  with  each   joint  actuated   by   coreless   DC   motors   via   distributed   control   with   controllable  compliance.   DARwIn-‐XOS   was   developed   through   collaboration   between   Virginia  Tech   and   the   University   of   Pennsylvania.   Using   a   computer   vision   system   and  accelerometer,   DARwIn-‐XOS   can   implement   human-‐like   gaits   while   navigating  obstacles  and  traverse  uneven  terrain  while  implementing  complex  behaviors  such  as  playing  soccer.    For  RoboCup  2012,  Team  VT  DARwIn  from  the  Humanoid  League  and  UPennalizers  from   the   Standard   Platform   League   are   continuing   to   team   up   together   to   form  Team   DARwIn   in   the   Humanoid   League.   With   Virginia   Tech’s   expertise   in  mechanical   design   and   University   of   Pennsylvania’s   expertise   in   software  engineering  and  strategy,  we  hope  to  replicate  our  success  in  the  highly  competitive  RoboCup  competition.    DARwIn  has  experience  in  both  the  Kid  and  Adult  size  leagues;  the  team  would  like  to  take  on  a  new  challenge  and  develop  within  the  TeenSize  League.  

 

 Fig.  1.  Family  of  DARwIn  Robots  

 Team  DARwIn  would  like  to  commit  to  participate  in  the  RoboCup  2012  Humanoid  League  competition.  Team  DARwIn  is  able  to  provide  students  who  participated  in  last  year’s  competition  to  serve  as  referees,  as  they  have  sufficient  knowledge  of  the  rules.    2  Research  The  DARwIn  family  serves  as  a  research  platform  used   for  studying  dynamic  gaits  and   walking   control   algorithms.   With   few   exceptions   (i.e.   the   Honda   ASIMO,   the  Sony  QRIO,  and  the  KAIST  HUBO  [3–7]),  most  legged  robots  today  walk  using  what  is  called  the  static  stability  criterion.  The  static  stability  criterion  is  an  approach  to  prevent  the  robot  from  falling  down  by  keeping  the  center  of  mass  of  its  body  over  the  support  polygon  by  adjusting  the  position  of  its  links  and  pose  of  its  body  very  slowly   to  minimize  dynamic  effects   [5].  Thus  at   any  given   instant   in   the  walk,   the  robot   could   “pause”   and   not   fall   over.   Static   stability   walking   is   generally   energy  inefficient  since  the  robot  must  constantly  adjust  its  pose  to  keep  the  center  of  mass  of  the  robot  over  its  support  polygon,  which  generally  requires  large  torques  at  the  joint   actuators   (similar   to   a   human   standing   still   with   one   foot   off   the   ground).  Humans   naturally   walk   dynamically   with   the   center   of   mass   rarely   inside   the  

support  polygon.  Thus  human  walking  can  be  considered  as  a  cycle  of  continuously  falling  and  catching  its  fall:  a  cycle  of  exchanging  potential  energy  and  kinetic  energy  of   the   system   like   the  motion   of   an   inverted   pendulum.  Humans   fall   forward   and  catch   themselves  with   the   swinging   foot  while   continuously   progressing   forward.  This   falling   motion   al-‐   lows   the   center   of   mass   to   continually   move   forward,  minimizing   the   energy   that   would   reduce   the  momentum.   The   lowered   potential  energy  from  this  forward  motion  is  then  increased  again  by  the  lifting  motion  of  the  supporting  leg.    One  natural  question  that  arises  when  examining  dynamic  walking  is  how  to  classify  the   stability   of   the   gait.   Dynamic   stability   is   commonly   measured   using   the   Zero  Moment  Point  (ZMP),  which  is  defined  as  the  point  where  the  influence  of  all  forces  acting   on   the  mechanism   can   be   replaced   by   one   single   force   without   a  moment  term  [8].  If  this  point  remains  in  the  support  polygon,  then  the  robot  can  have  some  control   over   the   motion   of   itself   by   applying   force   and/or   torque   to   the   ground.  Once  the  ZMP  moves  to  the  edge  of  the  foot,  the  robot  is  on  the  verge  of  stability  and  can  do  nothing  to  recover  without  extending  the  support  polygon  (planting  another  foot   or   arm).   Parameterized   gaits   can   be   optimized   using   the   ZMP   as   a   stability  criterion  or  stable  hyperbolic  gaits  can  be  generated  by  solving  the  ZMP  equation  for  a   path   of   the   center   of   mass.   Additionally,   the   ZMP   can   be   measured   directly   or  estimated   during   walking   to   give   the   robot   feedback   to   correct   and   control   its  walking.  DARwIn   is  developed  and  being  used   for   research  on  such  dynamic  gaits  and  control  strategies  for  stability  [5,  9].    3  Hardware  DARwIn-‐XOS  has  20  degrees  of  freedom  (six  in  each  leg,  three  in  each  arm,  and  two  in   the   head).   The   robot’s   links   are   fabricated   out   of   aluminum.   The   robot   uses  Robotis’  Dynamixel  RX-‐28M  motors   for   the  arm   joints,  RX-‐10  motors   for   the  neck  joints,  EX-‐106  motors   for   the  knee   joints,  and  RX-‐64  motors   for   the  remaining   leg  joints.  [10].  The  motors  operate  on  a  serial  TTY  network,  allowing  the  motors  to  be  daisy  chained  together.  Each  motor  has  its  own  built-‐in  optical  encoder  and  position  feedback   controller,   creating   distributed   control.   The   computers,   sensors,  electronics,  and  computer  ports  are  distributed  about  DARwIn’s  upper  torso.    

 Fig.  2.  Robot  height  dimensions  

 The  DARwIn-‐XOS  stands  101  cm  in  full  height.    The  body  consists  of  70  cm  legs,  a  16.4  cm  torso,  and  14  cm  head.    The  heights  fulfill  all  of  the  TeenSize  requirements  for  total  height,  leg  length,  and  head  size.    4  Electronics  DARwIn   XOS’s   electronic   system   provides   power   distribution,   communication  buses,  computing  platforms,  and  sensing  schemes  aimed  at  making  sense  of  a  salient  environment.  DARwIn’s  power  is  provided  by  two  8.2V  (nominal)   lithium  polymer  batteries   wired   in   series   providing   a   total   of   16.4V   to   the   joint   actuators   and  electronics.   These   batteries   provide   2.1   Ah,   which   gives   DARwIn   a   little   over   10  minutes  of  run  time.    

 Fig.  3.  Electronics  architecture  

 Computing   tasks   are   performed   on   a   Compulabs   Fit-‐PC2   computing   system   that  runs  the  Intel  Atom  Z530  CPU  with  1GB  of  onboard  RAM  and  built-‐in  WiFi.  The  PC  runs  off   the  main  battery  supply,  and  consumes  8W  at   full  CPU  usage.   In  addition,  the   computer   connects   to   both   a   Philips   SPC1300   camera   and   an   Arduino  microcontroller  board.    The   Arduino   board,   featuring   the   ATmega1280,   acts   as   the   communication   relay  between   the   Dynamixel   motors   and   the   fitPC.   The   microcontroller   also   provides  sensor   acquisition   and   processing.   Switches,   a   6   degree   of   freedom   Inertial  Measurement  Unit   (IMU),   and   foot   sensors  will   aid   in   correcting  gait   cycles   in   the  face  of  perturbations.  A  block  diagram   that  outlines   the   computing   relationship   is  shown  in  Fig.  2.    5  Software  The  software  architecture  for  the  robots  is  shown  in  Fig.  3.  This  architecture  is  novel  in  that  it  uses  Lua  as  a  common  development  platform.  Since  many  of  the  students  do   not   have   strong   programming   backgrounds,   this   development   platform   allows  them   to   participate  more   fully   on   the   team.   Low-‐level   interfaces   to   the   hardware  level   are   implemented   as   C   routines   callable   from   Lua.   These   routines   provide  access   to   the   camera   and   other   sensors   such   as   joint   encoders   and   the   IMU,   and  allow  the  higher-‐level  routines  to  modify  joint  angles  and  stiffnesses.  

 Fig.  4.  Software  architecture  

 The  Lua  routines  consist  of  a  variety  of  modules,  layered  hierarchically:  

– Sensor   Module   that   is   responsible   for   reading   joint   encoders,   IMU,   foot  sensors,  battery  status,  and  button  presses  on  the  robot.  

– Camera   Interface  to  the  video  camera  system,  including  setting  parameters,  switching  cameras,  and  reading  the  raw  YUYV  images.  

– Effector  Module  to  set  and  vary  motor  joints  and  parameters,  as  well  as  body  and  face  LED’s.  

– Vision   Uses   acquired   camera   images   to   deduce   presence   and   relative  location  of  the  ball,  goals,  lines,  and  other  robots.  

– World   Models   world   state   of   the   robot,   including   pose   and   altered   ball  location.  

– Game  StateMch  Game  state  machine  to  respond  to  Robocup  game  controller  and  referee  button  pushes.  

– Head   StateMch   Head   state  machine   to   implement   ball   tracking,   searching,  and  lookaround  behaviors.  

– Body   StateMch   Body   state   machine   to   switch   between   chasing,   ball  approach,  dribbling,  and  kicking  behaviors.  

– Keyframe   Keyframe   motion   generator   used   for   scripted   motions   such   as  getup  and  kick  motions.  

– Walk  Omnidirectional  locomotion  module.    In   order   to   simplify  development,   all   interprocess   communications   are  performed  by  passing  data  structures  between  the  various  modules,  as  well  as  between  robots.  [11]    6  Vision    

 Fig.  5.  Visualization  of  the  color  segmentation  

 

In  each  new  setting,  we  may  encounter  different  field  conditions  such  as  a  change  in  lighting  or  the  actual  color  hue  of  the  field  objects.  In  order  to  account  for  this,  we  log  a  series  of  images  that  are  then  used  to  train  a  lookup  table.  A  GUI  tool  enables  us   to   define   the   YCbCr   values   that   correspond   to   green,   yellow,   white,   etc.   Once  these  specific  values  are  selected  and  defined,   the  distribution  of   the  points   in   the  color  space  are  spread  out  and  generalized  to  account  for  a  greater  variation.  This  is  done  with  a  Gaussian  mixture  model  that  analyzes  the  probability  density  function  of  each  of   the  previously  defined  pixel  values.  The  boundaries  of   the  color   classes  are  then  expanded  according  to  Bayes  Theorem.  We  can  then  process  the  individual  pixels  of  the  new  images  by  matching  their  YCbCr  values  to  the  broadened  definition  of  the  values  in  the  lookup  table.    After  the  image  is  segmented  into  its  corresponding  color  classes  using  the  look-‐up  table,  the  segmentation  is  bitwise  OR-‐ed  in  4x4  blocks.  The  initial  object  hypotheses  for   the   ball   and   goal   posts   are   found   by   finding   connected   components   in   the  smaller,   bit   OR-‐ed,   image,   and   then   using   the   original   image   we   calculated   the  statistics  of  each  region.  Processing  the  bit  OR-‐ed  image  first  allowed  us  to  greatly  speed  up  the  computation  of  the  system.  The  bit  OR-‐ed  image  also  produced  the  set  of  points  that  are  used  in  our  line  detection  algorithm.    We  then  check  the  segmented  components  for  certain  attributes  like  size,  shape,  and  position   in   order   to   classify   objects,   such   as   the   ball   and   the   goal   posts.  We   also  compute   statistics   for   the   position   of   detected   objects   in   the   world   coordinate  system  using  the  inverse  kinematics  of  the  robot,  the  centroid,  and  the  bounding  box  to  further  filter  the  object  hypotheses.  Using  these  we,  are  able  to  track  the  ball  and  identify  the  existence  and  size  of  goal  posts  and  consequently  localize  our  position  on  the  field.  [11]    7  Prior  Performance  in  RoboCup  Team  DARwIn  has  been  competing  since  RoboCup  2007  led  by  Virginia  Tech,  and  in  recent  years  has  partnered  with  the  University  of  Pennsylvania.  RoboCup  2010  saw  the   team   play   a   respectable   fourth   place,   and   drove   the   team   to   improve   its  hardware   and   software.   With   the   introduction   of   the   DARwIn-‐OP   system,   Team  DARwIn  was  able  to  move  ahead  and  enthusiastically  took  the  first  place  trophy  in  the   kid-‐size   league   in   Istanbul.   With   this   momentum,   we   are   hoping   to   innovate  even  more  this  coming  year.    This   year,   the   team   has   pushed   forward   innovate   on   the   hardware   end   of   the  DARwIn   Open   Platform;   we   have   developed   a   robot   to   compete   in   the   TeenSize  competition.     We   use   our   own   software,   and   provide   its   base   at  https://github.com/UPenn-‐RoboCup/UPennalizers  for  the  research  community.    8  Conclusion  Building  on  previous  research  and  RoboCup  experience,  we  hope  to  push  DARwIn-‐  OP   its   highest   potential.   The   combination   of   hardware,   electronic,   and   software  expertise  from  both  Virginia  Tech  and  University  of  Pennsylvania  should  prove  to  be  

very  powerful  and  make  Team  DARwIn  a   formidable  opponent  at  RoboCup,  while  still   maintaining   its   primary   purpose   as   a   research   platform   for   studying   robot  humanoid  motions.    References

1. D. Hong. Biologically inspired locomotion strategies: Novel ground mobile robots at RoMeLa. 3rd International Conference on Ubiquitous Robots and Ambient Intelligence, Seoul, S. Korea, October 2006.

2. K. Muecke and D. Hong. Darwinʼs evolution: Development of a humanoid robot. Proc. IEEE/RSJ Int. Conf. on Intelligent Robots and Systems, pages 2574–2575, October-November 2007.

3. K. Hirai. The Development of Honda Humanoid Robot. IEEE Int. Conf. on Robotics and Automation, pages 1321–1326, 1998.

4. T. Ishida. A small biped entertainment robot sdr-4x ii. Proc. IEEE Symposium on Comp. Intelligence in Robotics and Automation, pages 1046–1051, July 2003.

5. J. Kim. On the Stable Dynamic Walking of Biped Humanoid Robots. PhD Thesis, Korea Advanced Institute of Science and Technology, Daejeon, South Korea, 2006. et al S. H. Collins. A three-dimensional passive-dynamic walking robot with two legs and knees. Int. Journal of Robotics Research, 20(2):607–615, 2001.

6. T. McGeer. Passive dynamic walking. Int. Journal of Robotics Research, 9(2):62– 82, April 1990. 8. M. Vukobratovic. Zero-moment point–thirty five years of its life. Int. Journal of Humanoid Robotics, 1(1), 2004.

7. Q. Huang, K. Yokoi, and et al S. Kajita. Planning walking patterns for a biped robot. IEEE Trans. on Robotics and Automation, 17(3):280–289, June 2001.

8. K. Kim, Y. Son, and P. Kim. Construction of small humanoids with a new joint actuator module. Proc. 2004 IEEE Int. Conf. on Robotics and Automation, pages 4510–4514, April 2004.

9. J.Brindza, A.Lee, A.Majumdar, B.Scharfman, A.Schneider, R.Shor,and D.Lee. UPennalizers RoboCup Standard Platform League Team Report 2009. Technical report, University of Pennsylvania, 2009.